| As NAND Flash continue to grow enormously both in capacity and popularity, the adoption of Flash in server as storage is accelerating because of its superiority in access latency and energy consumption. However, unbalanced I/O load across different storage servers incurs wear unbalance and durability as well as reliability degradation. In addition, unbalance wear is mainly caused by the uneven write access, the garbage collection of the server with serious wear will be more frequently, resulting in long time response, it will be the bottleneck of the system.Dynamic data redirection technique based on consistent hashing is proposed to balance the wear across the servers in distributed storage system based on flash. Firstly, the technique dynamically redirect hot data from server with high erase counts to the server with low erase counts, and redirect cold data from server with low erase counts to the server with high erase counts, to accelerate the wear of the server with low erase counts and relax the wear of the server with high erase counts. Secondly, the technique exploits the hot data with frequent update and the out-of-place update feature of flash by redirecting the updates from source server to the destination server instead of moving data from source server to the destination server. It can reduce the amount of data migration, the write amplification and extra-cost. Moreover, the normal access will not be affected by delaying the migration of the hot data.The wear-leveling technique of the data redirection dynamically based on consistent hashing is implemented in the distributed storage system with simulation of flash by flash-sim. Experimental results show that, using the wear-leveling technique can make the write access and wear become balanced, and delaying the migration of the hot data can reduce the data migration amount of nearly 40% and reduce the extra-cost. The technique can minimize the average wear of servers in the system. |
PDF Full Text Request